MiDShip

MiDShip is a multimodal dataset of synthetically generated ship cargo-hold structures for machine learning, engineering design, and optimization. It connects 120-parameter design vectors to full and mesh-ready CAD geometry, engineering drawings, drawing annotations, structural-element tables, calculated structural properties, and 25 local scantling constraints derived from the American Bureau of Shipping (ABS) Marine Vessel Rules.

The current release contains 12,753 retained designs spanning tankers, containerships, and bulk carriers. The SGLD-generated and repaired subsets are derived from selected random-design seeds and should be treated as related subsets rather than statistically independent samples.

Accompanying code: MiDShip on GitHub

Dataset at a glance

Subset	Attempted candidates	Retained designs	Fully feasible	Mean constraint violations
Random	6,050	6,020	0.0%	13.192
SGLD-generated	500	496	64.9%	0.409
Equation-informed repaired	6,254	6,237	79.4%	0.296

The SGLD-generated subset was produced by a surrogate-guided inverse-design experiment. The repaired subset was produced by applying equation-informed parameter corrections to selected random designs. Feasibility requires satisfying all 25 implemented constraints.

Directory structure

The three release subsets use the following layout:

MiDShip_Dataset/
├── Random_Structures/
│   ├── batch_0000/ ... batch_0005/       # CAD, bills of materials, and tables
│   ├── Dataset_Drawings/
│   │   └── batch_0000/ ... batch_0009/   # Drawings and annotations
│   ├── Test_Designs/                      # Small development examples
│   └── Test_Drawings/                     # Small development examples
├── Repaired_Structures/
│   ├── batch_0000/ ... batch_0004/       # CAD, bills of materials, and tables
│   └── Dataset_Drawings/
│       └── batch_0000/ ... batch_0009/   # Drawings and annotations
└── SGLD_Gen_Structures/
    ├── sgld_design_<index>.*             # CAD and bills of materials
    ├── sgld_*.csv                         # Subset-level tables
    └── Dataset_Drawings/                  # Drawings and annotations

Random and repaired structures are divided among batch directories to keep the very large collections manageable. The batches are storage partitions, not train/test splits, and their design indices are not necessarily contiguous. SGLD structures are stored directly in the subset directory.

Random_Structures/Test_Designs and Test_Drawings contain development examples and are not additional members of the 12,753-design research cohort.

Design names and index alignment

Each attempted candidate has a zero-based integer index. The subset-specific file stems are:

Subset	File stem
Random	random_test_design_<index>
Repaired	repaired_random_design_design_<index>
SGLD-generated	sgld_design_<index>

The repeated design in the repaired stem is part of the released filename and should not be removed when constructing paths.

The aggregate CSV tables do not contain a separate design-ID column. Their zero-based row number is the candidate index used in the filenames. Candidate tables retain all attempted rows, including candidates for which geometry or evaluation failed, so released file indices are intentionally non-contiguous. Use the subset's *_error_idx*.csv file and missing structural-property rows when constructing an aligned retained-design table.

What each retained design contains

For a file stem such as random_test_design_0, the structure package contains:

File	Contents
random_test_design_0.3dm	Native Rhino model of the full structure
random_test_design_0.igs	General IGES export of the full structure
random_test_design_0_MeshElements.igs	Mesh-ready IGES export; smaller stiffeners are represented by curves for later line-element meshing
random_test_design_0_Structural_Elements.csv	Structural-element bill of materials and geometric/section properties

The structural-element table has one row per modeled element and 33 fields. These include object identifiers linking the table to the CAD objects, location and direction, dimensions and thickness, element class and type, panel area/volume/centroid, and cross-sectional properties. Object IDs are embedded in both the full and mesh-ready CAD representations to preserve traceability between modalities.

The release provides mesh-ready geometry, not finite-element meshes or FEA results.

Drawing package

Each retained design has three engineering drawing views:

• Midship Section IWO of Web Frame
• Midship Section of Long. Structure
• Transverse Bulkhead

Each view is provided as a standard PDF and a corresponding _with_BBoxes.pdf with labeled component bounding boxes. Two CSV files provide the machine-readable annotation data:

File suffix	Contents
_Drawing_Annotations.csv	Component Object_ID, drawing-view name, and Rhino drawing-object ID
_Slice_Elements.csv	Slice position, structural-element link, drawing scale, and bounding boxes in model and drawing coordinates

This gives each design three plain PDFs, three annotated PDFs, one drawing annotation table, and one slice-element table.

Subset-level tables

The current release stores the aggregate tables at these locations:

Subset	Parameter table directory	Prefix
Random	Random_Structures/batch_0005/	random_test_design_
Repaired	Repaired_Structures/batch_0000/	repaired_random_design_
SGLD-generated	SGLD_Gen_Structures/	sgld_

The principal tables are:

Table suffix	Columns	Description
Parameters_All.csv or parameters_Updated.csv	120 columns	Principal dimensions, ship class, plate dimensions, and stiffener/girder spacing and sizing
Structural_Properties.csv	12 columns	Steel volume and weight, unit steel weight, center of gravity, longitudinal cross-sectional properties, and maximum bending moment
Constraint_Values.csv	25 columns	Calculated value for each implemented constraint
Constraint_Thresholds.csv	25 columns	Governing threshold paired with each calculated constraint value

The exact parameter filenames are:

• random_test_design_Parameters_All.csv
• repaired_random_design_Parameters_Updated.csv
• sgld_design_parameters_Updated.csv

The 120 parameters comprise 10 principal-characteristic and ship-class fields, 27 major-plate fields, and 83 girder/stiffener fields. Ship class is one-hot encoded through the cat tanker, cat container, and cat bulkcarrier columns. Class-specific features are zero when inactive.

Loading an aligned subset

The following example loads the random candidate tables, filters incomplete structural-property rows, and preserves the candidate index used by the CAD and drawing filenames:

from pathlib import Path

import pandas as pd


dataset_root = Path("MiDShip_Dataset")
table_dir = dataset_root / "Random_Structures" / "batch_0005"

# Each table row uses the same zero-based candidate index as the filenames.
parameters = pd.read_csv(
    table_dir / "random_test_design_Parameters_All.csv"
)
properties = pd.read_csv(
    table_dir / "random_test_design_Structural_Properties.csv"
)
constraint_values = pd.read_csv(
    table_dir / "random_test_design_Constraint_Values.csv"
)
constraint_thresholds = pd.read_csv(
    table_dir / "random_test_design_Constraint_Thresholds.csv"
)

# A complete structural-property row identifies a retained random design.
retained = properties.notna().all(axis=1)

parameters = parameters.loc[retained].rename_axis("design_index")
properties = properties.loc[retained].rename_axis("design_index")

# Every stored constraint uses the same value >= threshold convention.
constraint_margin = (
    constraint_values.loc[retained]
    - constraint_thresholds.loc[retained]
).rename_axis("design_index")

fully_feasible = constraint_margin.ge(0).all(axis=1)

For candidate index 0, the corresponding random CAD package can be found under Random_Structures/batch_*/random_test_design_0.*, and its drawings under Random_Structures/Dataset_Drawings/batch_*/random_test_design_0_*.

Constraint interpretation

For each design and constraint, the evaluator stores a calculated value and a governing threshold. A constraint is satisfied when:

constraint value >= constraint threshold

The sole maximum-spacing constraint is sign-transformed before storage so that it follows the same comparison. The 25 constraints cover selected plate thicknesses, stiffener and girder section moduli, structural depths, and member spacing.

These constraints are a research-oriented subset of applicable structural rules. They are not a complete ship-classification approval process and do not replace full-vessel loading, finite-element analysis, or review by a classification society.

Accompanying code

The MiDShip GitHub repository is the reproducibility companion to this data release. Place this dataset directory at the repository root as MiDShip_Dataset/; the maintained scripts use that relative path.

The principal code components are:

Code path	Purpose
Rhino_Macros/Parametric_Structure_V2.py	Defines the parametric cargo-hold geometry inside Rhino
Rhino_Macros/rhino_StructGen.py	Exports structures and structural-element tables
Rhino_Macros/rhino_2D_Drawing.py	Produces engineering drawings and annotations
equation_repair_pipeline/batched_structure_generation.py	Shared restartable Rhino generation worker
Rhino_Macros/Batched_Drawing_Generation.py	Restartable drawing-generation supervisor
tools/Parametric_Structure_Eval.py	Calculates structural properties and the 25 constraint pairs
tools/evaluate_midship_dataset.py	Re-evaluates and aligns released subset tables
tools/repair_parametric_designs.py	Applies equation-informed parameter repair rules
sgld_generation_pipeline/sgld_experiment.py	Trains surrogates and generates SGLD candidates

Every subset follows the same data flow:

1. generate or repair parameter vectors;
2. generate full and mesh-ready structures in Rhino;
3. generate engineering drawings in Rhino;
4. evaluate structural properties and the 25 constraints.

The numerical and machine-learning environment is defined by Autogluon_env.yml. CAD and drawing generation additionally require Rhino 8 for macOS and its Python scripting interface. The random, repaired, and SGLD entry points are documented in the GitHub repository's top-level README.md, with method-specific details in equation_repair_pipeline/README.md and sgld_generation_pipeline/README.md.

Design-space relationship

The accompanying code generates a shared t-SNE visualization comparing the random, SGLD-generated, and repaired parameter sets in one joint embedding. It provides a qualitative view of the distributions; t-SNE is not used as a quantitative measure of global design-space coverage.

Scope and limitations

MiDShip contains synthetic cargo-hold regions rather than complete or empirical vessels. It omits full-vessel arrangements, vessel-level loading, finite-element results, localized reinforcement, and other higher-fidelity details. The prescribed parameter ranges and generation procedures should not be used to infer the prevalence of configurations in operational ships.

Manuscript and citation

The dataset and generation methods are described in the accompanying manuscript, “MiDShip: Multimodal Dataset of Ship Cargo Hold Structures for Engineering Design.” A formal citation and archival publication link will be added when they are available.

License

The MiDShip dataset is released under the GNU General Public License version 3 (GNU GPL-3.0).

Disclaimer

Disclaimer: This research was funded by the American Bureau of Shipping (ABS). The opinions, findings, conclusions, technical approach, analysis, calculations, and recommendations expressed herein, including any use, interpretation, or derivation of ABS Rule requirements or formulas, are solely those of the author(s) and have not been reviewed, validated, or endorsed by ABS. Nothing in this paper may be relied upon as a statement or interpretation of the ABS Rules or as a substitute for the ABS Rules as published by ABS, which govern in all cases. ABS makes no representation or warranty as to the accuracy or fitness for any purpose of the material herein and assumes no liability arising from its use.